Multiverse, Qualcomm Optimize AI Models for Dragonfly AI200 and AI250 as Data Centers Chase Efficiency
Updated
Updated · Pulse 2.0 · Aug 5
Multiverse, Qualcomm Optimize AI Models for Dragonfly AI200 and AI250 as Data Centers Chase Efficiency
3 articles · Updated · Pulse 2.0 · Aug 5
Summary
Multiverse Computing and Qualcomm Technologies said they will tune AI models for Dragonfly AI200 and AI250 accelerators to cut compute, memory and power needs in data centers.
The effort targets operators struggling to scale AI workloads under electricity, cooling, memory-capacity and accelerator constraints, with the companies pitching more inference capacity from existing hardware.
March 2026 tests on Qualcomm’s Cloud AI100 Ultra gave a preview: a compressed open-source LLM ran up to 93% faster, with 44% higher throughput, 45% lower memory use and 21% lower power consumption.
A separate on-premises financial-document chatbot ran up to 35% faster and delivered 54% greater throughput, also trimming memory by 45% and power by 14% without losing accuracy.
The companies expect newer Dragonfly chips could do better still, but disclosed no performance targets or launch timeline for jointly optimized models.